YouTube Clip Extractor

Overview

This skill downloads YouTube videos, analyzes transcripts for compelling clip moments, extracts clips using ffmpeg, and generates platform-ready on-screen text and captions. It integrates with the existing caption and social content skills to deliver complete, publishable assets.

When to Use This Skill

•You have a YouTube URL and want to extract the best clips
•You want automated clip identification based on hook/coda criteria
•You need clips cut and ready for Descript or other editors
•You want on-screen text hooks and platform-specific captions for each clip

Do NOT use for:

•Full podcast production workflow (use podcast-production skill instead)
•Text-only social posts (use social-content-creation skill)
•Already-downloaded videos (skip to Phase 2)

Prerequisites

Required Tools (install via Homebrew)

bash

brew install yt-dlp ffmpeg

File Location

All downloads go to your project's transcript folder (customizable).

Structure:

code

transcripts/
├── {video_id}.mp4              # Full video (H.264 encoded)
├── {video_id}.en.vtt           # Timestamped subtitles
└── clips/
    └── {video_id}/
        ├── clip_01_{name}.mp4  # Individual clips
        ├── clip_02_{name}.mp4
        └── {video_id}_Clip_Assets.md  # Captions & hooks

The 4-Phase Workflow

Phase 1: Download Video & Transcript

Goal: Get video and subtitles from YouTube URL

Option A: Transcript-First (Recommended)

Download transcript first, identify clips, then download only needed segments. See Phase 1.5.

Option B: Full Download

If you need the entire video, use H.264 format for Descript compatibility (NOT AV1):

bash

# Download video in H.264 format (Descript-compatible)
yt-dlp -f "bestvideo[vcodec^=avc]+bestaudio[ext=m4a]/best[vcodec^=avc]" \
  --merge-output-format mp4 \
  -o "transcripts/{video_id}.mp4" \
  "YOUTUBE_URL"

# If H.264 unavailable, download best quality then re-encode:
yt-dlp -f "bestvideo+bestaudio" --merge-output-format mp4 \
  -o "transcripts/{video_id}_temp.mp4" \
  "YOUTUBE_URL"

# Re-encode to H.264 for Descript compatibility
ffmpeg -i "{video_id}_temp.mp4" -c:v libx264 -preset fast -crf 22 \
  -c:a aac -b:a 128k "{video_id}.mp4"

Step 2: Download Subtitles

bash

yt-dlp --write-auto-sub --sub-lang en --skip-download \
  -o "transcripts/{video_id}.%(ext)s" \
  "YOUTUBE_URL"

Phase 1 Output:

•{video_id}.mp4 - Full video (H.264)
•{video_id}.en.vtt - Timestamped subtitles

Phase 1.5: Convert VTT to Markdown (Critical for Context Efficiency)

Goal: Transform verbose VTT files into clean timestamped markdown for efficient analysis

Raw VTT files are extremely inefficient - they contain duplicate lines, millisecond timestamps, and formatting codes that consume massive context when reading. Always convert before analyzing.

Conversion Script

Use the VTT-to-markdown converter script:

bash

python3 scripts/vtt_to_markdown.py "transcripts/{video_id}.en.vtt" "transcripts/{video_id}.md"

This produces clean markdown with timestamps every ~30 seconds:

code

[0:00] First paragraph of dialogue here...

[0:30] Next paragraph continues here...

[1:00] And so on...

Workflow Optimization: Transcript-First Approach

For bandwidth efficiency, download transcript BEFORE the full video:

bash

# Step 1: Download transcript only
yt-dlp --write-auto-sub --sub-lang en --skip-download \
  -o "transcripts/{video_id}.%(ext)s" \
  "YOUTUBE_URL"

# Step 2: Convert to markdown
python3 scripts/vtt_to_markdown.py "transcripts/{video_id}.en.vtt" "transcripts/{video_id}.md"

# Step 3: Analyze transcript and identify clips (delegate to subagent)
# Step 4: Only download video if clips are worth extracting

Cleanup

After successful markdown conversion, delete the VTT file to keep things clean:

bash

rm "transcripts/{video_id}.en.vtt"

Phase 1.5 Output:

•{video_id}.md - Clean timestamped markdown transcript (VTT deleted)

Phase 2: Identify the Best Clip (Opus)

Goal: Find the single best clip with narrative arc - use Opus model for quality

IMPORTANT: Use Opus (not Haiku) for clip identification. This requires judgment about story structure, not just keyword matching.

Output Location

All clip suggestions go to:

code

studio/social-media/youtube-clips/{video_id}_clips.md

Clip Selection Criteria

The ideal clip has TWO essential elements:

•
Compelling Hook (first 3 seconds must grab attention)
- •Provocative statement or objection being raised
- •Surprising claim or counter-intuitive setup
- •Clear stakes or tension
•
Satisfying Arc or Reveal (complete thought, not just a quote)
- •Setup → Tension → Payoff
- •Can include multiple beats for richer storytelling
- •Ends with insight, punchline, or memorable line

Bonus: Multiple beats in the story (setup, twist, escalation, resolution)

Target length: 60 seconds or under

Timestamp Format Rules

CRITICAL: Only add timestamps at edit points (cuts)

For consecutive/continuous segments, use a single timestamp range:

markdown

**7:39 - 8:19** Holt tells the Connecticut homeschool story...

For edited clips with cuts, mark each segment separately:

markdown

| Timestamp | Speaker | Content |
|-----------|---------|---------|
| **6:47 - 7:04** | Donahue | "...what you get sooner or later is a very spoiled child" |
| **[CUT]** | | |
| **7:39 - 8:19** | Holt | Connecticut girl story → "...doesn't work out in practice" |

The [CUT] marker indicates removed material. Only use multiple timestamp rows when there's an actual edit.

Output Format for Clip Suggestions

markdown

# {Video Title} - Clip Suggestions

**Source:** {YouTube URL}
**Video ID:** {video_id}
**Total Duration:** {length}

---

## Best Clip: "{Descriptive Title}"

**Type:** [Continuous | Edited]
**Total Length:** ~{X} seconds

### Edit Sheet

| Timestamp | Speaker | Content |
|-----------|---------|---------|
| **{start} - {end}** | {Name} | "{Key quote or description}" |
| **[CUT]** | | _(if applicable)_ |
| **{start} - {end}** | {Name} | "{Key quote or description}" |

### Why This Works

- **Hook:** {What grabs attention in first 3 seconds}
- **Arc:** {Setup → Tension → Payoff summary}
- **Coda:** {How it ends / final memorable line}

### Suggested On-Screen Text

1. "{Hook option 1}"
2. "{Hook option 2}"
3. "{Hook option 3}"

---

## Alternate Clips (if applicable)

{Same format for runner-up options}

Narrative Snippets Method

Look for moments with story structure, not just good quotes:

Story Beats to Identify:

•Setup: Context that makes the payoff land
•Tension/Objection: Conflict, challenge, or stakes
•Turn: The moment things shift
•Payoff: Resolution, insight, or punchline

Example (John Holt clip):

•Setup (objection): Donahue voices fear: "what you get is a spoiled child"
•[CUT] - skip intermediate discussion
•Turn + Payoff: Holt's counter-example demolishes the concern

Scan Transcript For:

Objection-Response Pairs:

•Audience/host raises concern → Guest delivers knockout response
•These create natural tension and resolution

Inflection Points:

•"Then everything changed..."
•"I realized..."
•"That's when I knew..."

Vulnerability + Resolution:

•Personal stakes, failures → How they overcame

Surprising Data/Evidence:

•Counter-intuitive findings that challenge assumptions

Character in Action:

•Showing, not telling
•Doing, not describing
•Specific moments, not abstractions

Quality Tests (Pass 4/5):

• Stranger Test: Would someone with zero context care?
• Itch Test: Creates need to know more?
• Stakes Test: Clear why it matters?
• Tease Test: Hints without giving away?
• Emotion Test: Feel something in first 5 seconds?

Phase 2 Output Format:

Create analysis document with clip recommendations:

markdown

# {Video Title} - Clip Analysis

## Video Details
- **URL:** [YouTube URL]
- **Duration:** [Total length]
- **Speaker(s):** [Names]
- **Topic:** [Primary subject]

---

## Recommended Clips

### CLIP 1: "{Descriptive Name}"
**Timestamp:** `MM:SS - MM:SS` (XX seconds)
**Hook:** [First line or opening moment]
**Arc:** [Setup -> Middle -> Ending summary]
**Coda:** [How it ends / final line]

**Key Quotes:**
- "[Verbatim quote 1]"
- "[Verbatim quote 2]"
- "[Verbatim quote 3]"

**Quality Tests:** Stranger YES | Itch YES | Stakes YES | Tease YES | Emotion YES
**Why It Works:** [1-2 sentence rationale]
**Priority:** HIGH / MEDIUM / LOW

---

### CLIP 2: "{Descriptive Name}"
[Repeat structure...]

---

## Summary Table

| # | Clip Name | Timestamp | Length | Hook | Coda | Priority |
|---|-----------|-----------|--------|------|------|----------|
| 1 | [Name] | MM:SS-MM:SS | XXs | Strong | Strong | HIGH |
| 2 | [Name] | MM:SS-MM:SS | XXs | Strong | Good | HIGH |
| 3 | [Name] | MM:SS-MM:SS | XXs | Good | Strong | MEDIUM |

Phase 3: Cut Clips with FFmpeg

Goal: Extract approved clips as separate video files

Cutting Commands

Basic clip extraction (fast, uses keyframes):

bash

ffmpeg -i "{video_id}.mp4" -ss MM:SS -to MM:SS -c copy \
  "clips/{video_id}/clip_01_{name}.mp4"

Precise cutting with re-encoding (slower but frame-accurate):

bash

ffmpeg -ss MM:SS -i "{video_id}.mp4" -t DURATION \
  -c:v libx264 -preset fast -crf 22 -c:a aac -b:a 128k \
  "clips/{video_id}/clip_01_{name}.mp4"

Notes:

•-ss before -i = faster seeking (recommended)
•-c copy = no re-encoding (fast but may have keyframe issues)
•-c:v libx264 = re-encode to H.264 (slower but precise)
•Use H.264 output for Descript compatibility

Batch Cutting Example

bash

# Create clips directory
mkdir -p "transcripts/clips/{video_id}"

# Cut each clip
ffmpeg -i "{video_id}.mp4" -ss 06:59 -to 08:10 -c copy "clips/{video_id}/clip_01_key_insight.mp4"
ffmpeg -i "{video_id}.mp4" -ss 20:50 -to 21:54 -c copy "clips/{video_id}/clip_02_main_story.mp4"
ffmpeg -i "{video_id}.mp4" -ss 25:00 -to 26:10 -c copy "clips/{video_id}/clip_03_surprising_turn.mp4"

Phase 3 Output:

•Individual MP4 files for each clip
•All files in clips/{video_id}/ directory
•H.264 encoded for Descript compatibility

Phase 4: Generate On-Screen Text & Captions

Goal: Create platform-optimized hooks and captions for each clip

This phase uses the video-caption-creation skill methodology.

For Each Clip, Generate:

1. On-Screen Text Hook (3-5 options)

The text that appears in the first 3 seconds of the video. Must be:

•2-4 words maximum (mobile readable)
•Stops the scroll
•Passes McDonald's Test (accessible language)
•Complements (not duplicates) audio

Hook Categories:

•Polarizing: "Most people get this wrong"
•Counter-Intuitive: "The worst decision was the best"
•Direct Challenge: "Stop doing this immediately"
•Curiosity Gap: "Then everything changed..."

2. Platform-Specific Captions

Platform	On-Screen Text	Caption Style	Hashtags
Instagram	Same	Short, emoji OK, accessible	5-10
TikTok	Same	Short, emoji OK, accessible	3-5
YouTube Shorts	Same	Short, minimal emoji	3-5 + #Shorts
Facebook	Same	Slightly longer, conversational, NO external links	0-2

Facebook Difference: Caption can be longer and more conversational. NO hashtags or external links (kills reach).

3. Algorithm Optimization

Per the Triple Word Score system:

•Audio: Topic words spoken in first 10 seconds
•On-Screen Text: Reinforces (not competes with) audio
•Caption: Topic-relevant keywords in first sentence
•Hashtags: Broad -> Mid -> Specific -> Niche (10-12 total)

Phase 4 Output Format:

Create file: clips/{video_id}/{video_id}_CLIP_PACKAGE.md

markdown

# {Video Title} - Clip Package

## Source Video
- **URL:** [YouTube URL]
- **Title:** [Video title]
- **Duration:** [Total length]
- **Downloaded File:** `{video_id}.mp4`

---

## Context

[2-3 sentences explaining the backstory needed to understand the clip. Who is the speaker? What's their situation? What happened before/after the moments in the clip? This context ensures on-screen text and captions are coherent with the actual story.]

---

## Editing Instructions

**SEQUENCE (Rearranged from original - NOT linear):**

| Order | Timestamp | Speaker | Line |
|-------|-----------|---------|------|
| 1 | MM:SS-MM:SS | [Name] | "[Verbatim quote]" |
| 2 | MM:SS-MM:SS | [Name] | "[Verbatim quote]" |
| 3 | MM:SS-MM:SS | [Name] | "[Verbatim quote]" |

**OPTIONAL EXTENSION:**

| Order | Timestamp | Speaker | Line |
|-------|-----------|---------|------|
| 4 | MM:SS-MM:SS | [Name] | "[Verbatim quote]" |

---

## On-Screen Text Hook Options

1. **[Hook text]** - [Category]
2. **[Hook text]** - [Category]
3. **[Hook text]** - [Category]
4. **[Hook text]** - [Category]
5. **[Hook text]** - [Category]
6. **[Hook text]** - [Category]
7. **[Hook text]** - [Category]
8. **[Hook text]** - [Category]
9. **[Hook text]** - [Category]
10. **[Hook text]** - [Category]

---

## Platform Captions

### TikTok / Instagram Reels / YouTube Shorts
[Caption text]

[Hashtags: 3-5]

---

### Facebook
[Longer caption, conversational, NO hashtags]

---

### LinkedIn
[Professional tone caption]

[Hashtags: 3-5]

On-Screen Text Hook Categories

•Story Setup - Provides context that makes the clip make sense (e.g., "First day on the job")
•Polarizing - Bold statement that divides opinion (e.g., "Most experts are wrong")
•Contrast - Juxtaposition that creates tension (e.g., "Top performer. Zero passion.")
•Curiosity Gap - Teases without revealing (e.g., "What happened next changed everything")
•Story Tease - Hints at narrative arc (e.g., "She quit three days later")
•Pattern Interrupt - Subverts expectations (e.g., "This isn't what it looks like")

Context-Caption Coherence

Critical: On-screen text and captions must be coherent with the actual story in the transcript. Before writing hooks:

•Understand the full context (who, what, when, why)
•Identify what viewers need to know for the clip to make sense
•Choose hooks that accurately represent the story
•Avoid hooks that would confuse viewers when they hear the audio

Example: If the clip shows someone criticizing their previous job, but they were actually recounting their first week before they loved it, hooks like "First week was brutal" or "It gets better - trust me" provide necessary context that makes the story coherent.

Complete Workflow Example

bash

# PHASE 1: Download
yt-dlp -f "bestvideo[vcodec^=avc]+bestaudio" --merge-output-format mp4 \
  -o "transcripts/abc123def.mp4" \
  "https://www.youtube.com/watch?v=abc123def"

yt-dlp --write-auto-sub --sub-lang en --skip-download \
  -o "transcripts/abc123def.%(ext)s" \
  "https://www.youtube.com/watch?v=abc123def"

# PHASE 2: Analyze transcript (manual review)
# Read VTT file, identify clips using criteria above

# PHASE 3: Cut clips
mkdir -p "transcripts/clips/abc123def"
ffmpeg -i "abc123def.mp4" -ss 06:59 -to 08:10 \
  -c:v libx264 -preset fast -crf 22 -c:a aac \
  "clips/abc123def/clip_01_key_insight.mp4"

# PHASE 4: Generate assets (create markdown file with hooks/captions)

Related Skills

•Upstream: youtube-downloader
•Enhanced by: video-caption-creation
•Feeds into: short-form-video, social-content-creation